At Meta, efficiency at hyperscale is a constant challenge. With over 3 billion users, even a 0.1% performance regression can consume massive amounts of additional power. To tackle this, Meta's Capacity Efficiency Program has developed a unified AI agent platform that automates both finding and fixing performance issues across the infrastructure. By encoding the domain expertise of senior engineers into reusable, composable skills, these agents save hundreds of megawatts of power and compress investigation times from hours to minutes. This article explores how this system works and its impact on Meta's efficiency journey.

The Two Sides of Efficiency at Hyperscale

Meta views capacity efficiency as a two-front battle: offense and defense. Each requires distinct strategies but shares a common goal—reducing power consumption without compromising performance.

The Offensive Approach: Proactive Optimization

On the offensive side, engineers actively search for opportunities to make existing systems more efficient. This involves analyzing code, identifying redundant operations, and deploying optimizations. Historically, this process relied heavily on manual expertise, creating a bottleneck. Now, AI agents can automatically profile performance, suggest improvements, and even generate ready-to-review pull requests. This accelerates the cycle from discovery to deployment, allowing the team to scale win delivery without proportionally increasing headcount.

The Defensive Approach: Regression Detection with FBDetect

Defensively, Meta uses FBDetect, an in-house regression detection tool. It monitors production resource usage and catches thousands of regressions every week. When a regression occurs, it must be root-caused to a specific pull request and mitigated quickly to prevent wasted power from compounding across the fleet. Previously, this investigation could take up to ten hours of manual work—time that could be better spent innovating. AI agents now automate much of this diagnosis, compressing it to roughly 30 minutes.

The AI Agent Platform: Encoding Expertise into Reusable Skills

The heart of this transformation is a unified platform that combines standardized tool interfaces with encoded domain expertise. Senior efficiency engineers have distilled their knowledge into modular skills that can be reused and composed by AI agents. These skills enable agents to autonomously investigate issues across both offense and defense domains. The platform acts as a single interface, allowing agents to interact with various tools and databases without fragmentation. This design ensures that as new expertise is gained, it can be easily integrated into the system, continuously expanding its capabilities.

Measurable Impact: From Hours to Minutes, Megawatts Saved

The results speak for themselves. The AI agent platform has already recovered hundreds of megawatts of power—enough to power hundreds of thousands of American homes for a year. On the defensive side, agent-assisted resolution of regressions has significantly reduced the time wasted on compounding inefficiencies. Offensively, AI-driven opportunity resolution is expanding to more product areas each half, handling a growing volume of wins that engineers would never have time to address manually. Together, these capabilities allow the Capacity Efficiency Program to grow its power savings without scaling the team at the same rate.

The Road Ahead: A Self-Sustaining Efficiency Engine

Meta's ultimate goal is a self-sustaining efficiency engine where AI handles the long tail of performance issues. The current platform is a major step in that direction. As the system learns from more deployments and feedback, it will become increasingly capable of identifying and fixing issues autonomously. This vision frees engineers from repetitive investigation tasks, enabling them to focus on innovating new products. With continuous improvements, the AI agent platform promises to keep Meta's infrastructure both efficient and scalable for years to come.